Distinguishing clinically diagnosed psychiatric patients from screen-positive college students: An interpretable...

The psychiatric burden among college students has escalated substantially. However, traditional campus-based screening programs remain hampered by high false-positive rates and limited psychiatric referral resources. This study developed an interpretable machine learning framework to distinguish screen-positive students from university-aged patients with clinically diagnosed psychiatric disorders using SCL-90 item-level responses.

A clinical-reference classification cohort was constructed by integrating a campus screening sample with a clinical psychiatric cohort, with both cohorts restricted using the same high-score SCL-90 eligibility criterion. Feature selection was conducted utilizing LASSO and Boruta algorithms. Seven machine learning models were evaluated, with the optimal model interpreted via an Ensemble SHAP approach to elucidate feature contributions and algorithmic logic.

The core cohort comprised 411 screen-positive students and 2706 patients with clinically diagnosed psychiatric disorders. The XGBoost model achieved the numerically highest discriminative performance, with an area under the curve (AUC) of 0.897 (95% CI: 0.866-0.924). SHAP analysis identified nervousness, suicidal thoughts, and self-blame as the principal positive contributors to diagnosed psychiatric reference-cohort membership. Conversely, uneasy with opposite sex and fear of fainting in public were inversely associated with reference-cohort membership. Using 15 core items, the College Psychiatric Referral Predictor (CPRP) was developed to visualize the model-estimated likelihood of belonging to the clinically diagnosed reference cohort and provide individual-level SHAP explanations.

This framework supports clinical-reference classification between screen-positive students and clinically diagnosed psychiatric patients. The CPRP may assist post-screening prioritization in campus mental health settings, although prospective deployment evaluation is required before routine implementation.